CN106021374A - Underlay recall method and device for query result - Google Patents
Underlay recall method and device for query result Download PDFInfo
- Publication number
- CN106021374A CN106021374A CN201610309835.4A CN201610309835A CN106021374A CN 106021374 A CN106021374 A CN 106021374A CN 201610309835 A CN201610309835 A CN 201610309835A CN 106021374 A CN106021374 A CN 106021374A
- Authority
- CN
- China
- Prior art keywords
- query
- resource
- feature
- target
- formula
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2457—Query processing with adaptation to user needs
- G06F16/24578—Query processing with adaptation to user needs using ranking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The embodiment of the invention discloses an underlay recall method and an underlay recall device for a query result. The underlay recall method for the query result comprises the steps of acquiring query resources associated with a target query formula from a resource library according to the target query formula input by a search user; acquiring comparison score features of the query resources, wherein the comparison score features comprise basic related features; inputting the comparison score features of the query resources into a pre-trained sorting model, acquiring a correlation score value corresponding to each query resource output by the sorting model, wherein the sorting model is a Gbrank model; and according to the correlation score values, sorting the query resources, and selecting the set number of target resources as an underlay recall result of the target query formula according to a sorting result. According to the technical scheme of the method and device, by adopting the Gbrank model, the traditional underlay recall method for the query result is optimized, and the correlation between the recalled target resources and the target query formula is enhanced.
Description
Technical field
The present embodiments relate to the information processing technology, method recalled by the bottom particularly relating to a kind of Query Result
And device.
Background technology
Divide the online retrieving System Back-end of commercial search engine (such as Baidu, Google, the product such as searching well)
Two logic sub-modules of module are recalled for accurate order module and resource.Resource is recalled module and is responsible for climbing from network
Worm crawl and Data integration and database construction resource collection (if the resource collection of Webpage search is webpage collection, picture searching
Resource collection is pictures etc.) in recall the subset of resources relevant to query formulation, accurate order module be responsible for by
Resource is recalled the subset of resources that module recalls and is ranked up from high to low the most certainly according to the degree of association with query formulation
Determine finally to be presented to user search effect.Resource recall module recall result determine accurate order module sequence
Resource collection, the effect of accurate order module is produced remote-effects.Resource is recalled module and is recalled resource phase
The degree high effect on accurate order module in pass can produce the impact of forward.
Traditional ordering strategy is generally the simple sort methods such as bucket sort, is typically based on minority (typically,
3-5 ties up) simple basis correlative character (such as text relevant etc.) is ranked up, and strategy is the ofest poor quality.
The shortcoming of prior art is: first, and the basic correlative character that traditional sort method participates in recalling is less,
In long query formulation, recall effects performance is poor；Secondly, bucket sort model needs manual analysis difference basis phase
Degree of association contrast between closing property feature and resource dependencies, and increase basis correlative character is required for every time
Repeating the contrast between the correlative character of each basis, increase and decrease basis correlative character is the most convenient, expansible
Property is poor；Again, bucket sort model determines bucket according to the degree of association of basis correlative character with resource dependencies
Sequentially, and use basic correlative character the most, come basic correlative character below and resource arranged
The disturbance degree of sequence is the least, the most once comes basic correlative character above inverse to the judgement of resource degree of association
Sequence, comes basic correlative character below and cannot be carried out correcting, it is impossible to given play to each basis correlative character
Differentiation effect to resource.
Summary of the invention
In view of this, the embodiment of the present invention provides the bottom of a kind of Query Result to recall method and apparatus, with excellent
Select the bottom recall technique of existing target resource, improve bottom and recall the target query of result and user's input
Degree of association between formula.
In first aspect, method recalled by the bottom embodiments providing a kind of Query Result, including:
According to the target query formula of search user's input, obtain from resources bank and associate with described target query formula
Query resource；
Obtaining the alignment score feature of each described query resource, wherein, described alignment score feature includes basis
Correlated characteristic；
The described alignment score feature of each described query resource is inputted to the order models of training in advance, obtains
Take what described order models exported, the relevance score value corresponding with each described query resource, wherein, described
Order models is that gradient rises sequence (Gradient Boosting Rank, Gbrank) model；
According to described relevance score value, each described query resource is ranked up, and selects according to ranking results
The target resource taking setting number recalls result as the bottom corresponding with described target query formula.
In second aspect, the embodiment of the present invention additionally provides the bottom of a kind of Query Result and recalls device, bag
Include:
Query resource acquisition module, for the target query formula according to search user's input, obtains from resources bank
Take the query resource associated with described target query formula；
Alignment score feature acquisition module, for obtaining the alignment score feature of each described query resource, its
In, described alignment score feature includes basis correlated characteristic；
Relevance score value output module, for inputting the described alignment score feature of each described query resource
To the order models of training in advance, obtain the output of described order models, corresponding with each described query resource
Relevance score value, wherein, described order models is Gbrank model；
Result-generation module recalled by bottom, for according to described relevance score value, to each described query resource
Be ranked up, and according to ranking results choose set number target resource as with described target query formula pair
Result recalled by the bottom answered.
The embodiment of the present invention is recalled in outcome procedure at acquisition bottom, uses Gbrank model to carry out resource row
Sequence, the method using machine learning, the relation between training data learning feature and dependency are given a mark,
It is compared to bucket sort model and needs the problem of the degree of association between manual analysis different characteristic and dependency, joint
Save substantial amounts of manpower and time, more convenient.And traditional bucket sort model increases comparison every time
Scoring feature is required for repeating the contrast between each alignment score feature, and Gbrank model can be complete
Automatization is carried out, it is only necessary to add in training data newly-increased alignment score feature and again training pattern be
Can.There is priority in the alignment score feature used due to bucket sort model, uses multiple alignment score
During feature, the priority of alignment score feature is the lowest, the least to the disturbance degree of ranking results, it is impossible to have given play to
The differentiation effect to resource of each alignment score feature.And Gbrank model considers each feature to resource
Discrimination, is avoided that the defect of above bucket sort model.And once come feature above to degree of association
Judging backward, other features cannot be carried out correcting.Optimize the bottom recall technique of existing Query Result,
Being adjusted easily and optimize, flexibility ratio is higher, it is possible to be greatly improved target query resource defeated with user
Degree of association between the target query formula entered.
Accompanying drawing explanation
Fig. 1 is the flow chart that method recalled by the bottom of a kind of Query Result of first embodiment of the invention；
Fig. 2 is the flow chart that method recalled by the bottom of a kind of Query Result of second embodiment of the invention；
Fig. 3 is the flow chart that method recalled by the bottom of a kind of Query Result of third embodiment of the invention；
Fig. 4 is the flow chart that method recalled by the bottom of a kind of Query Result of fourth embodiment of the invention；
Fig. 5 is the structure chart that device recalled by the bottom of a kind of Query Result of first embodiment of the invention.
Detailed description of the invention
In order to make the object, technical solutions and advantages of the present invention clearer, below in conjunction with the accompanying drawings to the present invention
Specific embodiment is described in further detail.It is understood that specific embodiment described herein is only
It is only used for explaining the present invention, rather than limitation of the invention.
It also should be noted that, for the ease of describing, accompanying drawing illustrate only portion related to the present invention
Divide rather than full content.It should be mentioned that, some show before being discussed in greater detail exemplary embodiment
Example embodiment is described as process or the method described as flow chart.Although flow chart is by operations
(or step) be described as order process, but many of which operation can by concurrently, concomitantly or
Person implements simultaneously.Additionally, the order of operations can be rearranged.The most described place
Reason can be terminated, it is also possible to have the additional step being not included in accompanying drawing.Described process can be right
Should be in method, function, code, subroutine, subprogram etc..
First embodiment
The flow chart of method recalled by the bottom of a kind of Query Result that Fig. 1 provides for first embodiment of the invention,
The method of the present embodiment can be recalled device by the bottom of Query Result and be performed, this device can by hardware and/
Or the mode of software realizes, and typically can be integrated in for recalling relevant to the target query formula of user's input
In the server of target query resource.Technical scheme provided herein can formulate application by individualized feature
In all vertical classes and general search engine system.
The method of the present embodiment specifically includes:
110, according to the target query formula of search user's input, obtain and described target query formula from resources bank
The query resource of association.
In the present embodiment, the target query formula information of search user's input carries search-type, wherein,
Described search-type can be that user is selected by the option of search-type, it is also possible to is, target query
Formula itself carries search-type, and exemplarily, search-type includes webpage, picture, news, mhkc etc..
Such as, target query formula is " Beijing enlightening spy's layout ", and target query formula itself carries search-type and " schemes ".
Accordingly, the query resource associated with described target query formula can be specifically webpage collection, pictures or regard
Frequency collection waits the subset of resources associated with query formulation.
120, obtaining the alignment score feature of each described query resource, wherein, described alignment score feature includes
Basis correlated characteristic.
As it was previously stated, query resource is marked by the technical scheme of the present embodiment by use Gbrank model,
And be ranked up obtaining bottom according to appraisal result and recall result, and Gbrank model needs defeated when application
Enter one or more features of query resource to be marked, calculated by model and ultimately generate a score value.
Accordingly, in the present embodiment, described alignment score feature specifically refers to query resource needs input extremely
Described Gbrank model is to complete the feature finally marked.
Described basis correlated characteristic specifically refers to can be used to directly weigh between query resource and target query formula
The feature of dependency, typically, described basis correlated characteristic may include that target query formula and described inquiry
The matching degree feature of the webpage that resource is corresponding.
Preferably, in order to improve the final accuracy improving relativity measurement value, described alignment score further
Feature can also include: quality control feature, and/or click feature.
Specifically, described quality control characteristic may include that resource graded features, and resource area stepping
Feature.Exemplary, when described query resource is picture, resource graded features can include looking into described
Asking the website graded features that picture corresponding to resource is corresponding, further, website graded features can be specifically
(such as, the site level that clicking rate based on website, turnover rate, reliability assessment etc. determine marks
80 points or 90 are graded) or grade point (such as, A level or B level etc.).Resource area stepping feature
Can be the resolution of picture, size and/or the pixel value etc. that obtain of attribute according to picture.
Exemplarily, described click feature includes: precisely click feature and general click feature.
Concrete, accurate click feature can be to be analyzed by the data obtaining user behavior monitoring,
In acquisition user behavior data, all users are when inputting current goal query formulation, each target of corresponding click
The data of resource.Such as, accurate click feature is concrete it may be that correspond to the current of current goal query formulation
The user click data summation of target resource, the user of all target resources accounting for current goal query formulation represents
The ratio of data summation.Such as, picture and the correspondence that all users searching for model ice ice finally click on is recorded
Click total amount, then statistics picture represent total amount, and then the click total amount calculating each picture accounts for exhibition
The ratio of the click total amount that existing total amount is corresponding.It is for instance possible to use Wilson's space law obtains precisely clicks on spy
Levy.Wherein, represent total amount to obtain from the daily record of search engine.
General click feature can be that target query formula cutting generates participle collection, and then by user behavior is supervised
Survey the data obtained to be analyzed, obtain all users in user behavior data and input current goal query formulation
Corresponding each participle, the corresponding data clicking on each target resource.Such as, general click feature specifically may be used
To be, corresponding to the user click data summation of current goal resource of the current participle of current goal query formulation,
Account for the ratio of the user click data summation of all target resources of all participles corresponding to current goal query formulation
Example.Such as, current goal query formulation is " model ice ice force fawns ma's legend stage photo ", respectively general from off-line
The clicks that in click dictionary, four participles of retrieval are corresponding/represent number of times, calculate weighting based on participle weight respectively
With, then calculate, by click based on weighted sum/represent, the clicking rate that spotting out hits, as Wilson's can be used
Space law calculates the clicking rate that spotting out hits.Specifically, the building mode of offline dictionary is specifically as follows, by essence
The query formulation on schedule hit carries out participle, based on participle weight, the click of each resource/represent number is assigned to each participle
In unit, constituting the four-tuple of<term, obj, clc, pv>, i.e.<participle, query resource are clicked on participle, divide
Represent on word > four-tuple, be then keyword by participle and query resource, carrying out identical for keyword closed
And be added.General click feature is particularly suited for the less situation of click behavior of low frequency query formulation, it is possible to high
Frequently query formulation click feature is broken up in the click feature that participle is corresponding, and then is mapped in low frequency query formulation.
Preferably, when obtaining general click feature, in advance target query formula can be carried out pretreatment, only retain
Affect the important participle of target query formula and query resource dependency as current participle, for example, it is possible to go
After the function word in query formulation, obtain other participles remaining in target query formula as current participle.
It is to be appreciated that, in actual applications, basis correlated characteristic is relevant to the overall situation that bottom sorts, because of
This, can increase basis correlated characteristic according to the actual requirements, not limit at this.Such as, described basis
Correlated characteristic can also include: based on basic word number matching degree, and/or based on demand matching degree etc..Specifically,
Basic word number matching degree can be that the participle basic word number in hit territory accounts for territory always basic word number.
130, the described alignment score feature of each described query resource is inputted the order models to training in advance
In, obtain what described order models exported, the relevance score value corresponding with each described query resource, wherein,
Described order models is Gbrank model.
In this operation, by generally search for model the simple order models of bottom (as bucket sort, svm sequence
Deng) upgrade to more complicated Gbrank order models based on machine learning method, by Gbrank model
Select alignment score feature, by the method for machine learning, from training data learning alignment score feature with
Relation between dependency marking, uses the alignment score feature of various dimensions (such as basis correlative character, matter
Amount controlling feature, click feature etc.) carry out model training.When needing to increase new feature, and then in instruction
Practice and data add newly-increased alignment score feature training pattern again.
Wherein, first the training data of alignment score feature can draw from search in the way of using artificial mark
Some query formulation randomly drawed in the search record held up, these query formulation is submitted to Targeted Search Engine, so
After choose each or interval decimated search engine and return the front K of result, finally by professional, these are provided
Source is labeled according to the degree of association with query formulation.Wherein, K is positive integer.Such as, from search daily record
Randomly select a part of query formulation, allow the data assessment person by professional training provide query formulation and target resource
Dependency judge.The commonly scoring of fourth gear: poor, poor, preferably, good, in this, as training number
According to.
Can also excavate from daily record for corresponding search engine, search engine has substantial amounts of log recording
The behavior of user, can click on record by user and obtain training data.The Search Results that corresponding inquiry returns,
User can click on some webpage therein, it is assumed that what user priority was clicked on is the webpage more relevant with inquiry.One
As, user habit in from top to bottom browse Search Results, if user has skipped the webpage come below,
So come document below just more relevant than the document come above.
140, according to described relevance score value, each described query resource is ranked up, and according to sequence knot
Fruit is chosen the target resource of setting number and recalls result as the bottom corresponding with described target query formula.
In view of more with the query resource that target query formula is associated, therefore can choose according to ranking results and set
Fixed number purpose target resource, recalls result as the bottom corresponding with described target query formula.Wherein, set
Number can be configured according to practical situation, does not limits at this.Specifically, choose according to ranking results
Set number target resource recall result as the bottom corresponding with described target query formula it may be that according to
Ranking results, chooses the target setting number higher than certain threshold value of the relevance score value with target query formula
Resource, result recalled by the bottom corresponding with described target query formula.I.e. select the setting number that degree of association is higher
Target resource recall result as the bottom corresponding with described target query formula.
For example, user have input a query formulation " birthday card ", and search engine is based on this query formulation meeting
Have and recall certain Search Results, such as: " 1-website, website 5 ", need " birthday card " afterwards and
" 1-website, website 5 " be separately input in Gbrank model<birthday card, website 1>,<birthday card,
Website 2 > ... .., Gbrank model can export each website and the degree of association scoring being somebody's turn to do " birthday card ", and based on
1-website, website 5 is ranked up by similarity score result.
The embodiment of the present invention is recalled in outcome procedure at acquisition bottom, uses Gbrank model to be ranked up, makes
By the method for machine learning, the relation between training data learning feature and dependency are given a mark, compare
Then need the degree of association between manual analysis different characteristic and dependency in bucket sort model, save substantial amounts of
Manpower and time, more convenient.And traditional bucket sort model increases alignment score feature all every time
Need the contrast between repeating between each alignment score feature, and Gbrank model can enter with full automation
OK, it is only necessary in training data, add newly-increased alignment score feature training pattern again.Due to bucket
There is priority in the alignment score feature that order models uses, when using multiple alignment score feature, than
The lowest to the priority of scoring feature, the least to the disturbance degree of ranking results, it is impossible to have given play to each alignment score
The feature differentiation effect to resource.And Gbrank model considers each feature discrimination to resource, energy
Avoid the defect of above bucket sort model.And once come the judgement backward to degree of association of the feature above,
Other features cannot be carried out correcting.Optimize the bottom recall technique of existing Query Result, carry out easily
Adjusting and optimize, flexibility ratio is higher, it is possible to is greatly improved target query resource and looks into the target that user inputs
Degree of association between inquiry formula.
Second embodiment
Fig. 2 is the flow chart that method recalled by the bottom of a kind of Query Result of second embodiment of the invention.This reality
Execute example to be optimized based on the various embodiments described above, in the present embodiment, by the most excellent for basis correlated characteristic
Turn to: the matching degree feature of the webpage that described target query formula is corresponding with described query resource.
Accordingly, the basic correlated characteristic obtaining each described query resource is specifically optimized for: look into according to setting
Inquiry formula hierarchical rule, is decomposed into the participle collection of at least two level by described target query formula；Obtain one successively
Individual query resource, as current operation resource, obtains target web resource corresponding to described current operation resource
The single domain of the first quantity and the hybrid domain of the second quantity；Calculate the participle collection of each level and each single domain text respectively
And the object matching degree of each hybrid domain text, and using calculated described object matching degree as with described
The basic correlated characteristic that current operation resource is corresponding；Return and perform to obtain a query resource as current operation
The operation of resource, until completing the process to whole query resource.
Concrete, the method for the present embodiment includes:
110, according to the target query formula of search user's input, obtain and described target query formula from resources bank
The query resource of association.
121, according to setting query formulation hierarchical rule, described target query formula is decomposed at least two level
Participle collection.
In the present embodiment, setting query formulation hierarchical rule can be to use existing segmenter, according to existing point
The word segmentation regulation of word device carries out participle, and then according to the basic meaning of one's words, participle is divided level.Exemplarily,
After can be according to target query formula participle, each participle importance in target query formula, by target query formula
Being decomposed into core layering and adjust power layering, wherein, the participle collection of core layering includes the participle that must hit,
I.e. participate in the participle of correlation calculations marking, adjust the importance of the participle in power layering to be only second in core layering
The importance of participle, further, it is also possible to decomposite and target query formula disables layering, including mesh
Mark query formulation does not affect the word of the meaning of one's words.It should be noted that word is drawn by segmenter difference here that use
Divide also the most different.Usually, decomposition goal query formulation need to input target query formula carry out key word cutting,
Go the operation such as stop words, specifically can carry out participle according to individual character participle or according to the basic meaning of one's words.
In the present embodiment, set query formulation hierarchical rule and query formulation can also be carried out pretreatment, according to language
Meaning, linguistic context carry out participle, and then according to the part of speech of each participle, target query formula are decomposed at least two layer
The participle collection of level.Such as, in the participle corresponding to target query formula, nominal participle can be divided into
Core is layered, and verb participle is divided into tune power layering, is divided into by auxiliary word participle and disables layering etc..
122, one query resource of acquisition, as current operation resource, obtains described current operation resource pair successively
The single domain of the first quantity of the target web resource answered and the hybrid domain of the second quantity.
Wherein, the single domain and of the first quantity of target web resource corresponding to described current operation resource is obtained
The hybrid domain of two quantity is concrete it may be that according to setting web page resources resolution rules, to described target web
Resource carries out structuring parsing, generates single domain and the hybrid domain of the second quantity of the first quantity.For example, it is possible to
According to web page contents by target web resource resolution corresponding for current operation resource for including title field, subtitle
Territory, text field etc..It should be noted that owing to web page contents is different, structure is also not quite similar, so " the
One quantity ", " the second quantity " can be configured according to the actual requirements, do not limit at this.
In this operation, concrete can also obtain the mesh that described current operation resource is corresponding from inverted index
Mark the single domain of web page resources, and then the text attribute combining the participle of query formulation itself (such as participle length, is divided
Layer etc.) calculate basis correlative character.
123, participle collection and each single domain text and the object matching of each hybrid domain text of each level are calculated respectively
Degree, and calculated described object matching degree is relevant as the basis corresponding to described current operation resource
Feature.
One of the present embodiment preferred embodiment in, described object matching degree may include that based on length
The matching degree of degree, and/or matching degree based on weight.Specifically, matching degree based on length is it may be that root
Account for territory total length accounting according to the layering length in hit territory and calculate a matching degree；Matching degree based on weight it may be that
Participle weight sum according to hit territory accounts for matching degree accounting calculation of accounting calculation of the total weight in territory and once mates
Degree.
Exemplarily, target web resource structures is used to distinguish after dissolving analysis if basis correlative character increases
Six single domains and a hybrid domain, core layering, tune power that query formulation decomposites according to the importance of participle are divided
Layer and disable point word sets such as layering, and based on length with single domain text/hybrid domain text/query formulation text
Matching degree, and matching degree based on weight.Then basis correlative character amounts to 3*7*2=42 dimension.It is compared to
Traditional bottom method for retrieving, adds the basic correlative character participating in more recalling so that correspond to
The Query Result of long query formulation is more accurate.
124, the process to whole query resource has been judged whether；If so, 130 are performed；Otherwise, return
Perform 122.
130, the described alignment score feature of each described query resource is inputted the order models to training in advance
In, obtain what described order models exported, the relevance score value corresponding with each described query resource, wherein,
Described order models is Gbrank model.
140, according to described relevance score value, each described query resource is ranked up, and according to sequence knot
Fruit is chosen the target resource of setting number and recalls result as the bottom corresponding with described target query formula.
The technical scheme that the present embodiment is provided, by being decomposed into dividing of at least two level by target query formula
Word set, calculates participle collection and each single domain text and the object matching degree of each hybrid domain text of each level respectively,
As the basic correlated characteristic corresponding with described current operation resource.It is compared to traditional bottom and recalls method
Only use the basic correlative character of 3 to 5 dimensions, the basic correlative character that the technical program is used, energy
The Query Result enough making bottom recall is more accurate, and the method using machine learning, and flexibility ratio is higher.
3rd embodiment
Fig. 3 is the flow chart that method recalled by the bottom of a kind of Query Result of third embodiment of the invention.This reality
Execute example to be optimized based on the various embodiments described above, in the present embodiment, by described by each described inquiry money
The described alignment score feature in source inputs to the order models of training in advance, obtains the output of described order models
, the relevance score value corresponding with each described query resource is optimized for: obtain described 3rd quantity successively
Described query resource is as parallel processing resources, and according to the described sub-line journey set up, parallel acquisition is each described
The alignment score feature of parallel processing resources, and obtain the phase corresponding with each described parallel processing query resource
Closing property score value；Wherein, obtain a target parallel process the alignment score feature of resource and obtain described
Target parallel processes the operation of relevance score value corresponding to resource order in same thread and performs.
Concrete, the method for the present embodiment includes:
110, according to the target query formula of search user's input, obtain and described target query formula from resources bank
The query resource of association.
120, obtaining the alignment score feature of each described query resource, wherein, described alignment score feature includes
Basis correlated characteristic.
131, the sub-line journey of the 3rd quantity is set up.
Usually, at least one thread can be created program starts when, first create thread and become to serve as theme
Journey, usual sub-line journey, in order to complete certain task, is parallel to other threads of main thread.Considering would generally
There is the substantial amounts of query resource relevant to target query formula, for time-consuming raising efficiency, can use
The way of parallel processing, sets up the sub-line journey with the 3rd quantity.Wherein, the 3rd quantity can be that people is with the most whole
Number, concrete numerical value can select according to the actual requirements, not limit at this.In the present embodiment, may be used
It is set with the quantity according to the query resource corresponding with query formulation and/or default process time.
132, the described query resource of described 3rd quantity is obtained successively as parallel processing resources, and according to building
Vertical described sub-line journey, the parallel alignment score feature obtaining each described parallel processing resources, and according to than
Scoring feature is obtained the relevance score value corresponding with each described parallel processing query resource；Wherein, obtain
One target parallel processes the alignment score feature of resource and obtains described target also according to alignment score feature
Row processes the operation of relevance score value corresponding to resource order in same thread and performs.
In this operation, obtain the described query resource of described 3rd quantity successively as parallel processing resources,
I.e. can process multiple queries resource simultaneously, and set up multiple sub-line journey, according to the described sub-line journey set up,
Owing to the processing procedure of each sub-line journey is independent, obtain a target parallel and process the alignment score of resource
Feature and obtain described target parallel and process the operation of relevance score value corresponding to resource in same thread
Order performs, and does not affects mutually, is not to wait between multiple queries resource.Therefore, it is possible to obtain each institute parallel
State the alignment score feature of parallel processing resources, and obtain the output of described order models with each described parallel
Process the relevance score value that query resource is corresponding.Preferably.Described order models is Gbrank model.
140, according to described relevance score value, each described query resource is ranked up, and according to sequence knot
Fruit is chosen the target resource of setting number and recalls result as the bottom corresponding with described target query formula.
The technical scheme that the present embodiment is provided, obtains the query resource of described 3rd quantity successively as parallel
Process resource, and according to the sub-line journey set up, the parallel alignment score spy obtaining each described parallel processing resources
Levy, and obtain the relevance score value corresponding with each parallel processing query resource, and then according to described relevant
Property score value, each query resource is ranked up, and according to ranking results choose set number target resource
Result is recalled as the bottom corresponding with target query formula.Due to Gbrank model when alignment score only with treat
The alignment score feature calculating target query resource is correlated with, and multithreads computing therefore can be used each described
The relevance score value that query resource is corresponding, is ranked up each described query resource the most again, saves a large amount of
The operation time, ensureing while rate of precision, improving the efficiency that the bottom of Query Result is recalled, optimization is called together
Return performance.
4th embodiment
Fig. 4 is the flow chart that method recalled by the bottom of a kind of Query Result of third embodiment of the invention.This reality
Execute example to be optimized based on above-described embodiment, in the present embodiment, by described according to search user's input
Target query formula, from resources bank, obtain the query resource that associates with described target query formula be optimized for: be true
The type of fixed described target query formula；If described target query formula is short query formulation, then obtain and described mesh
The standard comparison feature that mark query formulation is corresponding；According to described standard comparison feature, obtain and institute from resources bank
State the query resource of target query formula association.
Concrete, the method for the present embodiment includes:
111, according to the target query formula of search user's input, the type of described target query formula is determined.
In the present embodiment, the target query formula of user's input can include long query formulation and short query formulation two kinds
Type, wherein, " length ", " short " can judge according to default query formulation judgment rule, such as, can
With the character length according to query formulation, or the participle quantity etc. in query formulation judges.
112, judge that described target query formula is short query formulation, if so, perform 113；Otherwise, 120 are performed.
Method for cutting is: selects the most one-dimensional basis correlative character, takes eigenvalue maximum topN
Query resource, N can arrange bigger herein, such as 100,000 ranks
113, the standard comparison feature corresponding with described target query formula is obtained.
Specifically, standard comparison feature can select corresponding with described target query formula the most at least one
Wiki plinth correlative character.Such as, the text relevant feature etc. extracted according to poly-justice.
114, according to described standard comparison feature, obtain and looking into that described target query formula associates from resources bank
Ask resource.
For short target query formula, due to the whole query resource quantity obtained according to described standard comparison feature
More, the query resource associated with described target query formula obtained from resources bank can be cut in advance
Disconnected, take the top n query resource that standard comparison eigenvalue is maximum, in order to ensure to recall the accuracy of resource,
N can arrange bigger herein, such as 100,000 ranks.
120, obtaining the alignment score feature of each described query resource, wherein, described alignment score feature includes
Basis correlated characteristic.
130, the described alignment score feature of each described query resource is inputted the order models to training in advance
In, obtain what described order models exported, the relevance score value corresponding with each described query resource, wherein,
Described order models is Gbrank model.
140, according to described relevance score value, each described query resource is ranked up, and according to sequence knot
Fruit is chosen the target resource of setting number and recalls result as the bottom corresponding with described target query formula.
What what the technical scheme that the present embodiment is provided can not only solve was corresponding to long target query formula is accurate
Resource recalls the problem of deficiency, and the number of the target query resource to be sorted in view of short target query formula
The target query resource that amount is corresponding much larger than long target query formula, the fewer appearance of short-tail target query formula simultaneously
Basis the inaccurate problem of correlation calculations, the technical program when determining that target query formula is short query formulation,
Obtain the standard comparison feature corresponding with described target query formula, according to described standard comparison feature, from resource
Storehouse obtains the query resource associated with described target query formula, i.e. certain can be utilized to calculate short query formulation
Basis correlative character blocks in advance the most accurately, while ensureing rate of precision, improves further and looks into
Ask the efficiency that the bottom of result is recalled, optimize and recall performance.
Searching system typically has multiple order module, and the sequence of usual bottom is simple, and the sequence on upper strata is complicated.
On the basis of the various embodiments described above, according to described relevance score value, each described query resource is carried out
Sequence, and choose the target resource of setting number as corresponding with described target query formula according to ranking results
After result recalled by bottom, preferably also include: described bottom is recalled result transmission and precisely sorts to upper strata
Model, recalls result according to described bottom carry out described target resource so that carrying out the accurate order models in upper strata
Sequence, and the ranking results of described target resource is fed back to user；Wherein, the accurate order models in upper strata is
Gbrank model.Use the technical program, use the bottom of Gbrank model realization target resource to recall, protect
Demonstrate,prove bottom and recalled the rate of precision of result, and then it is corresponding to use the accurate order models in upper strata that bottom is recalled result
Being ranked up of target resource.Due to bottom recall the high accurancy and precision of result be upper strata precisely sort accurate
Degree is had laid a good foundation, and drastically increase between target resource to the query formulation that user inputs is relevant
Property.
Further, in order to the consumption of central processing unit is greatly reduced, keep bottom to recall the essence of result simultaneously
Parasexuality, reduces response time.Preferably, a tree of the tree of the order models that bottom is recalled is accurate less than upper strata
A tree of the tree of order models.It is understood that " bottom ", " upper strata " are the orders processed for data
For, for distinguishing the operation performed by different application scene order models, make the statement of order models more
Add clear, not the restriction to order models.
5th embodiment
Fig. 5 is the structure chart that device recalled by the bottom of a kind of Query Result of fifth embodiment of the invention.Such as Fig. 5
Shown in, described device includes query resource acquisition module 510, alignment score feature acquisition module 520, is correlated with
Result-generation module 540 recalled by property score value output module 530 and bottom.
Wherein, query resource acquisition module 510, for the target query formula according to search user's input, from money
Storehouse, source obtains the query resource associated with described target query formula；
Alignment score feature acquisition module 520, for obtaining the alignment score feature of each described query resource, its
In, described alignment score feature includes basis correlated characteristic；
Relevance score value output module 530, for by defeated for the described alignment score feature of each described query resource
Enter to the order models of training in advance, obtain the output of described order models, with each described query resource pair
The relevance score value answered, wherein, described order models is Gbrank model；
Result-generation module 540 recalled by bottom, for according to described relevance score value, to each described inquiry money
Source is ranked up, and according to ranking results choose set number target resource as with described target query formula
Result recalled by corresponding bottom.
The embodiment of the present invention is recalled in outcome procedure at acquisition bottom, uses Gbrank model to be ranked up, makes
By the method for machine learning, the relation between training data learning feature and dependency are given a mark, compare
Then need the degree of association between manual analysis different characteristic and dependency in bucket sort model, save substantial amounts of
Manpower and time, more convenient.And traditional bucket sort model increases alignment score feature all every time
Need the contrast between repeating between each alignment score feature, and Gbrank model can enter with full automation
OK, it is only necessary in training data, add newly-increased alignment score feature training pattern again.Due to bucket
There is priority in the alignment score feature that order models uses, when using multiple alignment score feature, than
The lowest to the priority of scoring feature, the least to the disturbance degree of ranking results, it is impossible to have given play to each alignment score
The feature differentiation effect to resource.And Gbrank model considers each feature discrimination to resource, energy
Avoid the defect of above bucket sort model.And once come the judgement backward to degree of association of the feature above,
Other features cannot be carried out correcting.Optimize the bottom recall technique of existing Query Result, carry out easily
Adjusting and optimize, flexibility ratio is higher, it is possible to is greatly improved target query resource and looks into the target that user inputs
Degree of association between inquiry formula.
On the basis of above-described embodiment, specifically, described alignment score feature can also include: quality control
Feature processed, and/or click feature.
On the basis of the various embodiments described above, described quality control characteristic may include that resource graded features,
And resource area stepping feature.
On the basis of the various embodiments described above, described click feature specifically may include that accurate click feature with
And general click feature.
On the basis of the various embodiments described above, described basis correlated characteristic includes: described target query formula and institute
State the matching degree feature of webpage corresponding to query resource.Alignment score feature acquisition module specifically for: be used for
According to setting query formulation hierarchical rule, described target query formula is decomposed into the participle collection of at least two level；
For obtaining a query resource successively as current operation resource, obtain described current operation resource corresponding
The single domain of the first quantity of target web resource and the hybrid domain of the second quantity；Calculate the participle of each level respectively
Collection and each single domain text and the object matching degree of each hybrid domain text, and by calculated described target
Degree of joining is as the basic correlated characteristic corresponding with described current operation resource；Return and perform to obtain an inquiry money
Source is as the operation of current operation resource, until completing the process to whole query resource.
On the basis of the various embodiments described above, described object matching degree may include that matching degree based on length,
And/or matching degree of based on weight.
On the basis of the various embodiments described above, described basis correlated characteristic can also include: based on basic word number
Flux matched degree, and/or demand matching degree.
On the basis of the various embodiments described above, described relevance score value output module specifically may be used for: depends on
The described query resource of described 3rd quantity of secondary acquisition is as parallel processing resources, and according to the described son set up
Thread, the parallel alignment score feature obtaining each described parallel processing resources, and according to alignment score feature
Obtain the relevance score value corresponding with each described parallel processing query resource；Wherein, a target is obtained also
Row processes the alignment score feature of resource and obtains described target parallel process resource according to alignment score feature
The operation of corresponding relevance score value order in same thread performs.
On the basis of the various embodiments described above, described query resource acquisition module specifically for: determine described mesh
The type of mark query formulation；If described target query formula is short query formulation, then obtain and described target query formula
Corresponding standard comparison feature；According to described standard comparison feature, obtain from resources bank and look into described target
The query resource of inquiry formula association.
The bottom of the Query Result that the embodiment of the present invention is provided is recalled device and be can be used for performing the embodiment of the present invention
Method recalled by the bottom of the Query Result provided, and possesses corresponding functional module, it is achieved identical useful effect
Really.
Obviously, it will be understood by those skilled in the art that each module or each step of the above-mentioned present invention can be led to
Cross server as above to implement.Alternatively, the embodiment of the present invention can be able to perform with computer installation
Program realize, perform such that it is able to be stored in storing in device by processor, described journey
Sequence can be stored in a kind of computer-readable recording medium, and storage medium mentioned above can be read-only depositing
Reservoir, disk or CD etc.；Or they to be fabricated to respectively each integrated circuit modules, or by them
In multiple modules or step be fabricated to single integrated circuit module and realize.So, the present invention is not restricted to
The combination of any specific hardware and software.
The foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, for this area skill
For art personnel, the present invention can have various change and change.All institutes within spirit and principles of the present invention
Any modification, equivalent substitution and improvement etc. made, should be included within the scope of the present invention.
Claims (18)
1. method recalled by the bottom of a Query Result, it is characterised in that including:
According to the target query formula of search user's input, obtain from resources bank and associate with described target query formula
Query resource；
Obtaining the alignment score feature of each described query resource, wherein, described alignment score feature includes basis
Correlated characteristic；
The described alignment score feature of each described query resource is inputted to the order models of training in advance, obtains
Take what described order models exported, the relevance score value corresponding with each described query resource, wherein, described
Order models is Gbrank model；
According to described relevance score value, each described query resource is ranked up, and selects according to ranking results
The target resource taking setting number recalls result as the bottom corresponding with described target query formula.
Method the most according to claim 1, it is characterised in that described alignment score feature also includes:
Quality control feature, and/or click feature.
Method the most according to claim 2, it is characterised in that described quality control characteristic includes: money
Source graded features, and resource area stepping feature.
Method the most according to claim 2, it is characterised in that described click feature includes: precisely point
Hit feature, and general click feature.
5. according to the arbitrary described method of claim 1-4, it is characterised in that described basis correlated characteristic bag
Include: the matching degree feature of the webpage that described target query formula is corresponding with described query resource；
The basic correlated characteristic obtaining each described query resource includes:
According to setting query formulation hierarchical rule, described target query formula is decomposed into the participle of at least two level
Collection；
One query resource of acquisition is as current operation resource successively, obtains described current operation resource corresponding
The single domain of the first quantity of target web resource and the hybrid domain of the second quantity；
Calculate participle collection and each single domain text and the object matching degree of each hybrid domain text of each level respectively,
And calculated described object matching degree is relevant special as the basis corresponding to described current operation resource
Levy；
Return the operation performing to obtain a query resource as current operation resource, until completing all looking into
Ask the process of resource.
Method the most according to claim 5, it is characterised in that described object matching degree includes: based on
The matching degree of length, and/or matching degree based on weight.
Method the most according to claim 5, it is characterised in that described basis correlated characteristic also includes:
Based on basic word quantity Matching degree, and/or based on demand matching degree.
Method the most according to claim 5, it is characterised in that the described institute by each described query resource
State alignment score feature to input to the order models of training in advance, obtain the output of described order models, with
The relevance score value that each described query resource is corresponding, specifically includes:
Set up the sub-line journey of the 3rd quantity；
Obtain the described query resource of described 3rd quantity successively as parallel processing resources, and according to setting up
Described sub-line journey, the parallel alignment score feature obtaining each described parallel processing resources, and comment according to comparison
Dtex is levied and is obtained the relevance score value corresponding with each described parallel processing query resource；
Wherein, obtain a target parallel and process the alignment score feature of resource, and special according to alignment score
The operation order in same thread levying relevance score value corresponding to acquisition described target parallel process resource is held
OK.
Method the most according to claim 1, it is characterised in that the described mesh according to search user's input
Mark query formulation, obtains the query resource associated with described target query formula from resources bank and specifically includes:
Determine the type of described target query formula；
If described target query formula is short query formulation, then obtain the standard ratio corresponding with described target query formula
To feature；
According to described standard comparison feature, from resources bank, obtain the inquiry money associated with described target query formula
Source.
10. device recalled by the bottom of a Query Result, it is characterised in that including:
Query resource acquisition module, for the target query formula according to search user's input, obtains from resources bank
Take the query resource associated with described target query formula；
Alignment score feature acquisition module, for obtaining the alignment score feature of each described query resource, its
In, described alignment score feature includes basis correlated characteristic；
Relevance score value output module, for inputting the described alignment score feature of each described query resource
To the order models of training in advance, obtain the output of described order models, corresponding with each described query resource
Relevance score value, wherein, described order models is Gbrank model；
Result-generation module recalled by bottom, for according to described relevance score value, to each described query resource
Be ranked up, and according to ranking results choose set number target resource as with described target query formula pair
Result recalled by the bottom answered.
11. devices according to claim 10, it is characterised in that described alignment score feature is also wrapped
Include: quality control feature, and/or click feature.
12. devices according to claim 11, it is characterised in that described quality control characteristic includes:
Resource graded features, and resource area stepping feature.
13. devices according to claim 11, it is characterised in that described click feature includes: precisely
Click feature and general click feature.
14. according to the arbitrary described device of claim 10-13, it is characterised in that the relevant spy in described basis
Levy and include: the matching degree feature of the webpage that described target query formula is corresponding with described query resource；
Alignment score feature acquisition module specifically for:
For according to setting query formulation hierarchical rule, described target query formula being decomposed at least two level
Participle collection；
For obtaining a query resource successively as current operation resource, obtain described current operation resource pair
The single domain of the first quantity of the target web resource answered and the hybrid domain of the second quantity；
Calculate participle collection and each single domain text and the object matching degree of each hybrid domain text of each level respectively,
And calculated described object matching degree is relevant special as the basis corresponding to described current operation resource
Levy；
Return the operation performing to obtain a query resource as current operation resource, until completing all looking into
Ask the process of resource.
15. devices according to claim 14, it is characterised in that described object matching degree includes: base
In the matching degree of length, and/or matching degree based on weight.
16. methods according to claim 14, it is characterised in that described basis correlated characteristic also wraps
Include: based on basic word quantity Matching degree, and/or demand matching degree.
17. devices according to claim 14, it is characterised in that described relevance score value output mould
Block specifically for:
Set up the sub-line journey of the 3rd quantity；
Obtain the described query resource of described 3rd quantity successively as parallel processing resources, and according to setting up
Described sub-line journey, the parallel alignment score feature obtaining each described parallel processing resources, and comment according to comparison
Dtex is levied and is obtained the relevance score value corresponding with each described parallel processing query resource；
Wherein, obtain a target parallel and process the alignment score feature of resource and according to alignment score feature
The operation order in same thread obtaining relevance score value corresponding to described target parallel process resource is held
OK.
18. devices according to claim 10, it is characterised in that described query resource acquisition module has
Body is used for:
Determine the type of described target query formula；
If described target query formula is short query formulation, then obtain the standard ratio corresponding with described target query formula
To feature；
According to described standard comparison feature, from resources bank, obtain the inquiry money associated with described target query formula
Source.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610309835.4A CN106021374A (en) | 2016-05-11 | 2016-05-11 | Underlay recall method and device for query result |
| US15/281,748 US10366093B2 (en) | 2016-05-11 | 2016-09-30 | Query result bottom retrieval method and apparatus |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610309835.4A CN106021374A (en) | 2016-05-11 | 2016-05-11 | Underlay recall method and device for query result |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106021374A true CN106021374A (en) | 2016-10-12 |
Family
ID=57100569
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610309835.4A Pending CN106021374A (en) | 2016-05-11 | 2016-05-11 | Underlay recall method and device for query result |
Country Status (2)
| Country | Link |
|---|---|
| US (1) | US10366093B2 (en) |
| CN (1) | CN106021374A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777282A (en) * | 2016-12-29 | 2017-05-31 | 百度在线网络技术（北京）有限公司 | The sort method and device of relevant search |
| CN107908783A (en) * | 2017-12-07 | 2018-04-13 | 百度在线网络技术（北京）有限公司 | Retrieve appraisal procedure, device, server and the storage medium of text relevant |
| CN108988382A (en) * | 2018-07-09 | 2018-12-11 | 内蒙古电力(集团)有限责任公司内蒙古电力科学研究院分公司 | The grid-connected monitoring and evaluation method and system of photo-voltaic power generation station |
| CN109684207A (en) * | 2018-12-14 | 2019-04-26 | 平安科技（深圳）有限公司 | Method, apparatus, electronic equipment and the storage medium of sequence of operation encapsulation |
| CN110275779A (en) * | 2019-06-20 | 2019-09-24 | 北京百度网讯科技有限公司 | A kind of resource acquiring method, device, equipment and storage medium |
| CN110750617A (en) * | 2018-07-06 | 2020-02-04 | 北京嘀嘀无限科技发展有限公司 | Method and system for determining relevance between input text and interest points |
| CN111221943A (en) * | 2020-01-13 | 2020-06-02 | 口口相传(北京)网络技术有限公司 | Query result matching degree calculation method and device |
| CN111813888A (en) * | 2019-04-12 | 2020-10-23 | 微软技术许可有限责任公司 | Training target model |
| CN112131491A (en) * | 2020-11-20 | 2020-12-25 | 震坤行网络技术(南京)有限公司 | Hierarchical ordering method, computing device and computer readable storage medium |
| CN115062184A (en) * | 2022-06-29 | 2022-09-16 | 四川长虹电器股份有限公司 | Film sequencing method in voice recall scene |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11875706B2 (en) | 2018-02-20 | 2024-01-16 | Pearson Education, Inc. | Systems and methods for automated machine learning model training quality control |
| US11449762B2 (en) | 2018-02-20 | 2022-09-20 | Pearson Education, Inc. | Real time development of auto scoring essay models for custom created prompts |
| WO2019237298A1 (en) | 2018-06-14 | 2019-12-19 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for on-demand services |
| CN109471926A (en) * | 2018-10-30 | 2019-03-15 | 广东原昇信息科技有限公司 | Intelligent word making method based on NLP and company information |
| CN111125086B (en) * | 2018-10-31 | 2023-02-24 | 北京国双科技有限公司 | Method, device, storage medium and processor for acquiring data resources |
| CN112579767B (en) * | 2019-09-29 | 2024-05-03 | 北京搜狗科技发展有限公司 | Search processing method and device for search processing |
| CN113495984A (en) * | 2020-03-20 | 2021-10-12 | 华为技术有限公司 | Statement retrieval method and related device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090248668A1 (en) * | 2008-03-31 | 2009-10-01 | Zhaohui Zheng | Learning Ranking Functions Incorporating Isotonic Regression For Information Retrieval And Ranking |
| CN102999538A (en) * | 2011-09-08 | 2013-03-27 | 富士通株式会社 | Character searching method and equipment |
| CN104077306A (en) * | 2013-03-28 | 2014-10-01 | 阿里巴巴集团控股有限公司 | Search engine result sequencing method and search engine result sequencing system |
| CN104850531A (en) * | 2014-02-19 | 2015-08-19 | 日本电气株式会社 | Method and device for establishing mathematical model |
| CN105224959A (en) * | 2015-11-02 | 2016-01-06 | 北京奇艺世纪科技有限公司 | The training method of order models and device |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2005103951A1 (en) * | 2004-04-23 | 2005-11-03 | Novauris Technologies Limited | Tree index based method for accessing automatic directory |
| US7643822B2 (en) * | 2004-09-30 | 2010-01-05 | Google Inc. | Method and system for processing queries initiated by users of mobile devices |
| JP2008538019A (en) * | 2005-01-31 | 2008-10-02 | ムスグローブ テクノロジー エンタープライジィーズ，エルエルシー | System and method for generating linked classification structures |
| JP4251652B2 (en) * | 2006-06-09 | 2009-04-08 | インターナショナル・ビジネス・マシーンズ・コーポレーション | SEARCH DEVICE, SEARCH PROGRAM, AND SEARCH METHOD |
| CN102591876A (en) * | 2011-01-14 | 2012-07-18 | 阿里巴巴集团控股有限公司 | Sequencing method and device of search results |
| US8713028B2 (en) * | 2011-11-17 | 2014-04-29 | Yahoo! Inc. | Related news articles |
| US9959482B2 (en) * | 2014-11-28 | 2018-05-01 | Canon Kabushiki Kaisha | Classifying method, storage medium, inspection method, and inspection apparatus |
| US20170004455A1 (en) * | 2015-06-30 | 2017-01-05 | Linkedin Corporation | Nonlinear featurization of decision trees for linear regression modeling |
| US10353504B2 (en) * | 2015-09-08 | 2019-07-16 | Samsung Electronics Co., Ltd. | User interface for computing devices equipped with pressure-sensitive displays |
| US10482091B2 (en) * | 2016-03-18 | 2019-11-19 | Oath Inc. | Computerized system and method for high-quality and high-ranking digital content discovery |
-
2016
- 2016-05-11 CN CN201610309835.4A patent/CN106021374A/en active Pending
- 2016-09-30 US US15/281,748 patent/US10366093B2/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090248668A1 (en) * | 2008-03-31 | 2009-10-01 | Zhaohui Zheng | Learning Ranking Functions Incorporating Isotonic Regression For Information Retrieval And Ranking |
| CN102999538A (en) * | 2011-09-08 | 2013-03-27 | 富士通株式会社 | Character searching method and equipment |
| CN104077306A (en) * | 2013-03-28 | 2014-10-01 | 阿里巴巴集团控股有限公司 | Search engine result sequencing method and search engine result sequencing system |
| CN104850531A (en) * | 2014-02-19 | 2015-08-19 | 日本电气株式会社 | Method and device for establishing mathematical model |
| CN105224959A (en) * | 2015-11-02 | 2016-01-06 | 北京奇艺世纪科技有限公司 | The training method of order models and device |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106777282B (en) * | 2016-12-29 | 2018-07-13 | 百度在线网络技术（北京）有限公司 | The sort method and device of relevant search |
| US10331685B2 (en) | 2016-12-29 | 2019-06-25 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method and apparatus for sorting related searches |
| CN106777282A (en) * | 2016-12-29 | 2017-05-31 | 百度在线网络技术（北京）有限公司 | The sort method and device of relevant search |
| CN107908783A (en) * | 2017-12-07 | 2018-04-13 | 百度在线网络技术（北京）有限公司 | Retrieve appraisal procedure, device, server and the storage medium of text relevant |
| CN110750617A (en) * | 2018-07-06 | 2020-02-04 | 北京嘀嘀无限科技发展有限公司 | Method and system for determining relevance between input text and interest points |
| CN108988382A (en) * | 2018-07-09 | 2018-12-11 | 内蒙古电力(集团)有限责任公司内蒙古电力科学研究院分公司 | The grid-connected monitoring and evaluation method and system of photo-voltaic power generation station |
| CN109684207A (en) * | 2018-12-14 | 2019-04-26 | 平安科技（深圳）有限公司 | Method, apparatus, electronic equipment and the storage medium of sequence of operation encapsulation |
| CN109684207B (en) * | 2018-12-14 | 2022-09-02 | 平安科技（深圳）有限公司 | Method and device for packaging operation sequence, electronic equipment and storage medium |
| CN111813888A (en) * | 2019-04-12 | 2020-10-23 | 微软技术许可有限责任公司 | Training target model |
| CN110275779B (en) * | 2019-06-20 | 2022-07-08 | 北京百度网讯科技有限公司 | Resource acquisition method, device, equipment and storage medium |
| CN110275779A (en) * | 2019-06-20 | 2019-09-24 | 北京百度网讯科技有限公司 | A kind of resource acquiring method, device, equipment and storage medium |
| CN111221943A (en) * | 2020-01-13 | 2020-06-02 | 口口相传(北京)网络技术有限公司 | Query result matching degree calculation method and device |
| CN111221943B (en) * | 2020-01-13 | 2023-08-08 | 口口相传(北京)网络技术有限公司 | Query result matching degree calculation method and device |
| CN112131491A (en) * | 2020-11-20 | 2020-12-25 | 震坤行网络技术(南京)有限公司 | Hierarchical ordering method, computing device and computer readable storage medium |
| CN112131491B (en) * | 2020-11-20 | 2021-02-26 | 震坤行网络技术(南京)有限公司 | Hierarchical ordering method, computing device and computer readable storage medium |
| CN115062184A (en) * | 2022-06-29 | 2022-09-16 | 四川长虹电器股份有限公司 | Film sequencing method in voice recall scene |
| CN115062184B (en) * | 2022-06-29 | 2024-05-28 | 四川长虹电器股份有限公司 | Film ordering method under voice recall scene |
Also Published As
| Publication number | Publication date |
|---|---|
| US20170329782A1 (en) | 2017-11-16 |
| US10366093B2 (en) | 2019-07-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106021374A (en) | Underlay recall method and device for query result | |
| West et al. | Mining missing hyperlinks from human navigation traces: A case study of Wikipedia | |
| US6775674B1 (en) | Auto completion of relationships between objects in a data model | |
| CN101630314B (en) | Semantic query expansion method based on domain knowledge | |
| RU2549121C2 (en) | Merging search results | |
| CN102737120B (en) | Personalized network learning resource recommendation method | |
| CN108717433A (en) | A kind of construction of knowledge base method and device of programming-oriented field question answering system | |
| CN107145496A (en) | The method for being matched image with content item based on keyword | |
| CN106339756A (en) | Training data generation method and device and searching method and device | |
| CN106709040A (en) | Application search method and server | |
| KR20040016799A (en) | Document retrieval system and question answering system | |
| CN105765573A (en) | Improvements in website traffic optimization | |
| KR20020075600A (en) | Image search system and method | |
| CN101419625B (en) | Deep web self-adapting crawling method based on minimum searchable mode | |
| CN105975596A (en) | Query expansion method and system of search engine | |
| CN103150369A (en) | Method and device for identifying cheat web-pages | |
| CN104462399B (en) | The processing method and processing device of search result | |
| CN110008309A (en) | A kind of short phrase picking method and device | |
| CN107562966A (en) | The optimization system and method based on intelligence learning for web page interlinkage retrieval ordering | |
| CN110378206A (en) | A kind of intelligence Audit System and method | |
| CN105760504A (en) | Resource retrieval method based on browser | |
| CN106815265A (en) | The searching method and device of judgement document | |
| CN104615621B (en) | Correlation treatment method and system in search | |
| CN109388690A (en) | Text searching method, inverted list generation method and system for text retrieval | |
| Shah et al. | An agent based personalized intelligent E-learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161012 |